WebMCP declarative API: what the check tests and how to fix it

Tracked, not scored. Passes if any crawled page has a form with a toolname attribute; otherwise skipped. The declarative API is still a proposal.

Agent Protocols · Tracked, not scored

How common is it?

1% of 1223 measured sites publish a passing result (observed adoption; this check does not change the score).

Latest audit of each site under rubric v3.1. A site with nothing published counts as not passing. A partial result does not count as a pass.

What this check tests

This is a tracked check: it does not change your score. It reads the HTML of every page the audit crawled (the homepage plus up to nine more) and looks at each <form>:

  • A form with a non-empty toolname attribute is recorded as a WebMCP tool declaration, with its page and tool name.
  • A form without one is counted as a plain task form only when it is not a search, login or newsletter form. Search is a form with role="search", inside a <search> element, with an input type="search", or a single text field named q, query or s. Login is any form with a password field. Newsletter is a single email field. Everything else is a task form.

One or more declarations is a pass, 100, and the message reads, for example, "1/4 task forms declare WebMCP tools" (declared forms are counted whatever their type). No declaration is a skip with score 0. There is no fail and no partial.

What it does not check: whether tooldescription is present, whether toolparamdescription or toolautosubmit are used, or whether the browser accepts the tool. Forms added by JavaScript after load are not seen. Script-registered tools are the job of WebMCP imperative.

Why it matters to agents

WebMCP is a proposal from the W3C Web Machine Learning Community Group, for pages to expose "tools" to agents built into the browser. The group's README says it is currently focused on the imperative JavaScript API, and that a declarative counterpart based on HTML forms "is being explored separately" and "is not part of the specification at this time". The declarative explainer lists the input-schema algorithm as still to be defined. Chrome documents the attributes toolname and tooldescription on a form, optional toolparamdescription on fields, and toolautosubmit to let the agent submit without the user clicking. Chrome says that removing either toolname or tooldescription unregisters the tool, which is why a form with only toolname, which this check accepts, may not be a working tool.

Support is early. Chrome and Edge list origin trials, and Google's agent-friendly site guide says WebMCP is in active development and may change. None of the sources fetched says that any assistant ranks, recommends or cites sites because they declare tools. The audit tracks it so the adoption rate is visible, which is why a skip counts as "not adopted" in the rate shown on this page.

How to fix it

Treat this as an experiment, and only for forms that do something an agent could usefully do, such as a booking or support request:

<form toolname="createSupportRequest"
      tooldescription="Submit a request to the support team"
      action="/support">
  <label for="subject">Subject</label>
  <input id="subject" name="subject" type="text"
         toolparamdescription="Short summary of the problem" required>
  <button type="submit">Send</button>
</form>
  • Always set both toolname and tooldescription. Without a toolparamdescription, Chrome falls back to the field's <label>, so labelled fields help (structured forms).
  • Leave out toolautosubmit for anything with consequences. Without it the browser focuses the submit button and the user decides. Chrome's security page recommends hints such as consequentialHint for high-stakes imperative tools and warns that prompt injection cannot be fully prevented.
  • Keep tools few and distinct; Chrome notes each one uses context window.
  • Test in a browser with the origin trial or the flag listed in the implementation status.

If you run a server-side integration instead, see MCP endpoint. Skip this check if you have no agent-facing tasks.

Examples from measured sites

Up to 10 of the highest-scoring sites in each group, from recent audits. Failing includes partial results.

Passing

Sources

Verified against these sources on

Does your site pass this check?

Check your own site →